Search research

Using a large language model to play the user is now standard in scalable evaluation. It has a repeatedly diagnosed failure: the simulated user is excessively cooperative, so a system under test can score by the sheer number of questions it asks rather than by making the user willing to speak. We answer with a disclosure gate conditioning information release on the companion agent's behaviour: its state is a ladder of five ordered gates, merged onto three observable depth layers. We specify, ablate, and audit it, and train a user simulator against that specification. Gating behaviour is learned from the training corpus's synthetic branch, while the real branch supplies how people speak and react; after training, the simulator need not be told at runtime which gate each item sits behind. The gate is a load-bearing component of the environment: on the English corpus of a published companion-agent benchmark (CompanionBench), once training no longer states per example which gate each item sits behind, the largest rank displacement across 12 systems under test exceeds the noise band set by re-running that environment under a new seed, while per-system scores show no detectable change. We state two acceptance criteria: a ranking must be order-preserving, and absolute scores must be scale-stable. Of the candidates we examine, only one passes both -- the simulator we release -- and its leaderboard correlates at 0.993 with the benchmark's original simulator. By contrast, prompting a frontier model as the simulator barely moves the ranking while shifting every score upward -- a shift invisible to anyone checking the ranking alone. The environment we specify is the one that benchmark already used. That publication describes the mechanism in about four hundred words, and we supply what it lacked: specification, ablations, human studies, negative controls, and downstream sensitivity analysis.

Recent advances in large reasoning models (LRMs) have shown that reinforcement learning with verifiable rewards (RLVR) can substantially improve reasoning in mathematics and code, where outcomes can be checked automatically. Extending this progress to open-ended and agentic tasks remains difficult because reliable rewards are harder to obtain and direct human supervision cannot keep pace with the scale and complexity of model-generated experience. This paper studies how LRMs can continue to improve as human supervision gradually recedes from the learning loop. We examine two connected dimensions of this problem. The reward axis traces the development from per-instance human judgments to reusable verifiers and rewards that operate even without human feedback. The experience axis examines how learning can progress from human-curated tasks and environments toward self-generated curricula, constructed environments, and autonomous co-evolution. We connect these dimensions through a five-level ladder from L0 to L4 that identifies which parts of the learning process remain under continued human control. Our analysis further highlights the risks introduced by increasingly autonomous rewards and experience generation, including reward hacking, feedback drift, curriculum collapse, and environment errors. Consequently, we also provide the evaluation around three complementary objects: policy capability, feedback fidelity, and experience quality. This analysis provides a structured account of current approaches to scaling LRMs beyond human supervision and the open problems involved in developing self-sustaining learning systems toward superintelligence. Furthermore, we maintain a continuously updated \href{https://github.com/visitworld123/Awesome-Scaling-LRM-Beyond-Human-Supervision}{GitHub repository} to track the latest advances.

Large language models (LLMs) are used as post hoc explainers of sequential decision-making policies, producing natural-language explanations of why an action was chosen. However, LLMs often generate plausible but incorrect statements, and no existing approach systematically tests whether such explanations are faithful to the underlying environment. Two classic software testing challenges stand in the way: there is no oracle for the correctness of an explanation, and the test inputs, natural language queries about a policy's behavior, lack the structure needed for systematic test case generation. We address both. Probabilistic model checking provides the test oracle, computing exact reference results against which LLM answers are graded automatically. A taxonomy of post hoc query categories structures the input space around the environment-level facts from which policy explanations are composed; test cases generated from it are prioritized by question-specific diagnostic difficulty scores. Across seven MDP environments, the testing separates three open-weight LLMs: a reasoning model passes 85% of test cases, a mid-size model 70%, and a 1B model falls below the random baseline, while prioritization surfaces significantly harder cases than random selection. Our results indicate how trustworthy LLM-generated explanations are in model-free settings, where the same LLMs are used but no oracle exists to verify them.

4D understanding and reasoning is a fundamental capability for embodied AI agents operating in dynamic physical environments. However, existing vision encoders are largely limited to static 2D images or 3D point clouds without temporal modeling, or to 2D videos that lack accurate geometric depth reasoning. Consequently, current approaches fail to jointly capture spatial structure and motion evolution in dynamic scenes. We present CL4D, the first foundational 4D vision encoder that directly operates on dynamic point clouds, trained with a contrastive learning objective to align spatio-temporal geometric representations with natural language descriptions. By learning a shared embedding space between text and 4D scene dynamics, CL4D enables zero-shot motion-to-text and text-to-motion retrieval in dynamic environments and serves as a foundational 4D vision encoder for downstream 4D vision-language tasks. Building on this encoder, we introduce 4DVLM, a 4D vision-language model that conditions language generation on dynamic geometric representations. 4DVLM is the first VLM designed to operate directly on 4D point clouds without relying on 2D images, 2D videos, or static 3D point clouds. We train CL4D and subsequently 4DVLM on a newly constructed dataset termed DynAction4D capturing diverse human motions across varying object interactions and scene environments. Extensive experiments across multiple 4D human action benchmarks demonstrate that CL4D achieves state-of-the-art performance, with improvements of approximately ~16.75% over prior methods. Furthermore, 4DVLM outperforms frontier video VLMs such as Gemini and GPT-5 even when these models are provided with RGB video sequences corresponding to the same scenes represented as 4D point clouds for 4DVLM.

Training and evaluating spatial reasoning in embodied agents requires diverse environments that are both geometrically faithful and semantically queryable. Synthetic simulators offer ground truth semantics but sacrifice realism; simulators based on reconstructions of real-world environments have realistic appearance but lack ground truth semantics by default. We propose using Semantic Radiance Fields (SRF) as simulators for spatial reasoning agents. SRFs are a representation that unifies these requirements by lifting 2D semantic segmentations from pretrained vision models into a 3D radiance field that jointly encodes geometry, appearance, and per-class semantic identity. The resulting fields are reconstructed from posed RGB captures of real scenes and support novel-view synthesis, semantic and free-space queries within a single grounded representation. This enables the efficient generation of diverse real-world environments to train and evaluate spatial reasoning models. As an example application, we outline an SRF-driven simulator for an orchard apple-reaching task, in which the radiance field supplies camera rendering, semantic ground truth, and occupancy queries to a physics engine.

Automating empirical research is a long-standing direction of AI. Recent automatic research (AutoResearch) agents bring this goal within reach, as modern LLMs show the capability to independently implement solutions and learn from the execution outcomes. Behind these gains, post-training (especially RL) plays a central role. In this paper, we identify a fundamental tension when scaling RL for these agents: the two components of every AutoResearch trajectory (agent generation and environment execution) scale in very different manners, since all generation shares compute through batching, while each execution occupies its exclusive sandbox and real machine time. As a result, the environment execution dominates the training cost and becomes the bottleneck as trajectories grow. To resolve this tension, we propose World Model RL (WMRL), which replaces environment execution with a world model to remove this bottleneck. Additionally, the world model can be imperfect, as its rewards are corrupted by bias and noise. Therefore, we further equip WMRL with two mitigations, Online Debiasing and Inverse-Variance Denoising, which offset the bias and suppress the noise respectively. Theoretically, we prove that both mitigations of WMRL strictly improve the convergence guarantee. Empirically, WMRL accelerates training by 3-4x on various tasks at different agent scales, while exceeding the performance of standard RL baselines. Moreover, our post-trained 4B and 9B agents outperform much larger open-weight agents of 48B and 120B on held-out benchmarks. Beyond AutoResearch, WMRL also transfers to post-training embodied VLA policies, which demonstrates the generalizability of our method.

Tool-using LLM agents are commonly trained and evaluated in environments where tool calls succeed reliably, yet deployed tools can fail transiently, persistently, or silently. Robust recovery therefore requires more than repeated retries: an agent may need to retry the same path, switch to an alternative, or recognize that no viable path remains. We present BENCH2ROBUST, a framework that converts failure-free tool-use benchmarks into controlled stochastic environments with scenario-controlled solvability, where episodes explicitly require retrying, switching, or stopping after available paths are exhausted. We use BENCH2ROBUST to study two complementary interventions: structured runtime recovery context through Bayesian Tool Memory (BTM), and curriculum-controlled reinforcement learning. Across 7 models from 4 families and two multi-turn benchmark families, tool failures produce a near-universal robustness gap. On held-out Retail tasks, BTM improves robustness by up to 16.8 percentage points without retraining, while RL learns complementary recovery behavior that remains beneficial without inference-time BTM. Combining the two reaches 40.8-45.5% under injection while preserving failure-free performance. These results suggest that robust tool use benefits from combining environment-specific recovery knowledge with learned recovery behavior.

In sequential decision making, an agent typically observes its environment and acts at every timestep. However, such active participation may not always be necessary; tasks such as brewing coffee include periods that are served equally well by letting the environment evolve without constant monitoring and control. During such periods, the agent could simply wait to conserve its resources, or redirect its attention to another task. We capitalize on these opportunities by training a "waiting policy" that decides where and how long to wait. This involves forgoing sensing to commit to a wait action, representing a deliberate pause for a set number of timesteps. We formalize "learning to wait" as minimizing the frequency of sensing and decision making without sacrificing task performance (e.g., the total amount of time to complete a task). To train a waiting policy, we propose an approach that employs reinforcement learning with lexicographically ordered objectives. In experiments across 4 discrete-state household tasks and 3 continuous-state environments, we show that our approach successfully learns waiting behaviors, and can adapt pre-trained policies to wait where appropriate. While different tasks permit different amounts of waiting without sacrificing task performance, our approach consistently finds solutions with significant waiting, sometimes waiting for over 50 percent of the task duration.

Long-horizon tool-using agents must reason over user goals, domain policies, tool calls, simulator state, and delayed verifiable rewards. Reinforcement learning (RL) is a natural fit for this setting, but multi-turn on-policy rollouts create long contexts, while model-specific attention layers may require custom masks and learned sink normalization. We present SINKFLEX-RL, a modular training system for RL in dual-control tool-use environments. The system combines a Gymnasium-compatible environment wrapper, a VERL-style rollout dataflow, group-relative policy optimization without a separate value model, and a sink-aware FlexAttention path designed to preserve model-specific sink scaling under causal and sliding-window masks. In a preliminary Tau2Bench retail run, validation reward (mean@1) rises from 0.25 early in training to $0.44$ later in the observed training window, while training-score and trajectory-reward proxies also trend upward. In a fixed-configuration memory benchmark, the optimized attention path reduces peak VRAM from 28.06GB to 22.52GB at 4096 tokens, a $19.7\%$ reduction, and runs the measured 8192-token configuration using $25.53$~GB where the eager baseline runs out of memory. These results illustrate the value of integrating environment interfaces, RL dataflow, and attention-kernel design for memory-feasible long-horizon agent training.

Advanced air mobility operations hold the potential to enhance and expand regional transportation of both people and goods in populated areas. However, hazardous flight conditions arising from interactions between wind and the built environment remain a significant challenge for aerial vehicles in urban settings. This work proposes a novel framework towards safe flight planning of aerial vehicles in windy urban environments. A learning-based surrogate model is trained to rapidly predict flow fields from readily available information such as building geometry and incident wind. This surrogate prediction is used to calculate a volumetric flight challenge scalar field based on critical flow parameters and proximity to structures. A safe, flow-informed flight trajectory is then identified through a cost-minimizing pathfinder. The complete system is demonstrated experimentally through flight tests of a micro aerial vehicle through a model urban geometry placed in a large fan-array wind tunnel. Comparing this approach to trajectories generated without knowledge of the wind field, we find the flow-informed approach reduces undesired vehicle displacement and improves flight stability. This work is among the first practical demonstrations of safe, wind-aware methodologies for advanced air mobility in urban environments.

Robot policy evaluation and deployment remain fragmented by model-specific software dependencies, data representations, and runtime interfaces, so that connecting N policies to M evaluation environments requires O(NM) separate integrations. We present XPolicyLab, a unified standard and open ecosystem that reduces this cost to O(N+M). XPolicyLab specifies common observation, action, and trajectory schemas together with a minimal adapter interface for observation updates, action prediction, batched execution, and episode reset, while a dependency-isolated client/server architecture separates policy inference from environment execution, so that each side retains its native software stack and may run locally or remotely. The ecosystem integrates 42 robot policies and standardizes their installation, debugging, serving, and evaluation workflows. Across these adapters, model-specific code varies by an order of magnitude while the environment-facing loop stays within a few lines of a fixed reference, confirming that the contract confines heterogeneity to the policy side. In a controlled study, conforming to the standard reduces the integration effort of a representative policy from over five hours to two hours, and packaged agent skills reduce it further to thirty minutes. The same adapters serve RoboTwin, RoboDojo simulation, and standardized real-robot evaluation through one interface. XPolicyLab is released as shared infrastructure for reproducible policy comparison and standardized deployment across simulation and physical platforms. Project website: https://xpolicylab.github.io/.

We present 360CityArena, a benchmark for evaluating the urban exploration capabilities of embodied agents within a photorealistic environment constructed from 360-degree videos. Existing outdoor benchmarks either lack sufficient photorealism or complexity, resulting in a considerable gap from real-world urban environments. 360CityArena is built on a realistic reconstruction of the Akihabara district in Tokyo, Japan, using 602 360-degree video segments covering 85 streets, and consists of 175 meticulously human-crafted tasks. It encompasses three task categories: Environment Understanding, Path Reasoning, and Spatial Reasoning, covering fundamental abilities required for urban exploration, such as localization, landmark search, path planning, and relational spatial reasoning, thereby enabling comprehensive evaluation in realistic urban scenes. Our evaluation using state-of-the-art LMM-based agents shows that even the strongest model, Gemini 2.5 Flash, performs far below human level (human: 77.3% vs. Gemini 2.5 Flash: 17.1%), revealing substantial challenges that remain in city-scale embodied navigation and reasoning. 360CityArena provides a necessary and challenging testbed for photorealistic urban-district navigation and spatial reasoning.

Delicate manipulation often involves transient and subtle collisions between a grasped object and the environment. While the human hand localizes these contacts effortlessly thanks to superior tactile sensitivity, robotic systems often lack the requisite resolution to acquire the information necessary for motion planning, resulting in clumsy manipulation or even task failure. Here, we propose transient extrinsic contact detection and ranging (TECDAR), a simple yet fast and efficient method for detecting and ranging extrinsic contact of grasped objects. Our design of gripper tips employs dynamic tactile sensing leveraging a single 2.5$\times$3 mm 6D inertial measurement unit. The sensor captures sub-millisecond tip deformations at a 7 kHz sampling rate, but operating on a data stream of only 84 KB/s. High bandwidth and compact data size enable the system to rapidly detect and localize contact between grasped objects and their surroundings. Specifically, fusing tactile data with robot pose via an extended Kalman filter enables fast and precise localization of extrinsic contact, reaching millimeter-level accuracy within 180 ms. Experimental results demonstrate that the system achieves an average localization accuracy of approximately 7,mm in both line-contact and point-contact localization tasks. Furthermore, this near-instantaneous localization enables the robot to rectify its trajectory on a millisecond scale, facilitating precise tool manipulation and enhanced perception of complex environments purely through tactile exploration and mapping. We envision such techniques advancing the future of robotics across domains requiring delicate manipulation, including precision assembly, surgical assistance, and autonomous exploration in touch-dominant environments. Project page: humitlab.github.io/TECDAR/

Visual perception of walkability varies substantially across individuals, reflecting differences in personal characteristics, experiences, and preferences. Existing studies, however, often reduce these diverse judgements to aggregated scores, implicitly assuming uniform perception, and commonly rely on vehicle-mounted street-view imagery that does not reflect the pedestrian's visual experience. This paper introduces a dataset of 29,870 walkability ratings from 1,196 respondents, linking sidewalk-view imagery across urban, suburban, and regional Australian environments with individual rater attributes, and proposes the first user-conditioned multimodal deep learning framework for walkability perception, fusing visual features with respondent-level representations. A viewpoint-comparison study shows that sidewalk-view images receive significantly higher walkability ratings than matched street-view images, indicating that imagery source is a substantive design decision in perception surveys. The user-conditioned model improves rank agreement with observed ratings by 65% over an image-only baseline (quadratic weighted kappa 0.47 vs. 0.29), demonstrating that who is evaluating an environment carries predictive indication beyond image content alone. These findings support moving from aggregated, observer-independent walkability scores toward models that represent diverse users, enabling more inclusive assessment of pedestrian environments.

To address the issue that existing dynamic multi-objective optimization algorithms mainly rely on individual migration or independent special point sampling after environmental changes, while failing to fully exploit the structural relationships among representative solutions, a Special Point Skeleton Reconstruction based Dynamic Multi-Objective Evolutionary Algorithm (SPSR-DMOEA) is proposed. First, the centroid, knee points, and extreme points are extracted from the Pareto optimal solution set of the current environment, and their positions in the new environment are adaptively predicted according to their movement velocities across consecutive environments. Subsequently, in the decision space, the centroid is connected with other anchor points, and a minimum spanning tree is constructed among the non-centroid anchor points, thereby establishing a prediction skeleton capable of describing the overall population structure. According to the lengths of the skeleton edges, the number of individuals allocated to each edge is determined proportionally. Candidate solutions are uniformly generated along each edge, and random orthogonal perturbations are introduced to expand the search region around the skeleton. Experimental results on the DF dynamic multi-objective benchmark suite demonstrate the effectiveness of the proposed method in dynamic tracking capability.

Service robots operate in household environments shared with humans, pets, and everyday objects, where they are highly susceptible to failures such as software crashes, hardware degradation, or unpredictable interactions. While roboticists strive to minimize failures, some remain inevitable, making it critical to mitigate their potential consequences for safe and reliable deployment. This paper introduces a novel safety formulation that evaluates both the probability of impactful interactions between robots and surrounding entities during failures, and the severity of their outcomes. By quantifying the impact of failures on different entities, our approach enables robots to make informed planning decisions that balance safety with task efficiency. To support systematic evaluation, we also present FailBench, a MuJoCo-based simulation framework for studying robot-environment interactions under diverse failure modes, including sensing issues and actuator malfunctions. Together, our safety formulation and FailBench provide a foundation for developing safer and more robust motion plans and learned policies in real-world household environments.

Recent advances in machine learning have enabled training of wireless foundation models, which aim to support tasks such as channel estimation, beam prediction, and localization based on wireless signals. Existing wireless foundation models typically pretrain on channel tensors using masked reconstruction over subcarriers, antennas, or time but ignore the physical characteristics of wireless propagation. In this work, we propose to instead use multipath propagation as the fundamental pretraining object. We present MultiPathFormer, an autoregressive foundation model that represents each transmitter-receiver link as an ordered sequence of continuous-valued path tokens and pretrains with next-path prediction. We introduce an Environmental RAG (retrieval-augmented generation) mechanism and a first-path codebook on top of the transformer backbone, leveraging environment knowledge to improve path statistics estimation like delay and power by up to 59%. MultiPathFormer pretrained on 27 environments transfers to unseen users and, after scenario-specific fine-tuning, outperforms training the corresponding models from scratch in new environments. Across downstream tasks, it outperforms SOTA channel-based foundation models, achieving 5.57 m mean localization error, 0.914 top-3 beam accuracy, 0.994 line-of-sight classification accuracy, and 0.561 channel estimation NMSE. These results show that path-level pretraining can learn reusable representations of wireless propagation.

As AI systems become increasingly integrated into diverse interfaces and applications, model-centric audits are insufficient to address risks arising from interactions among system components and deployment environments. System integration has long been central to software audits in safety-critical domains such as aerospace. However, its role in AI auditing remains underexplored. Scanning through 4,259 documents, we present a scoping review of AI audits that treat system integration as a core tenet of evaluation (n = 58). Using reflexive thematic analysis, we analyze their elements, actors, enablers, and constraints. We find that the corpus represents an emerging yet still fragmented form of AI auditing: few existing measures target integration-specific risks; large gaps remain in meeting traditional audit expectations; and access to necessary information and resources significantly influences audit design. Nonetheless, integration can be categorized across three sites (inter-component, system-environment, and multi-system), each serving the functions of risk exploration, risk determination, coordination, and procedural regularity. Deviating from other types of evaluations, these audits assess qualities specific to system integration, including compatibility, completeness, and oversight. This review calls on the AI community to prioritize system integration as a core strategy for addressing AI risk, and to develop audit practices capable of capturing failures across components, environments, and systems beyond the reach of component-level evaluation.

Biological visual systems can perceive depth from monocular vision flow, continuously integrating temporal visual cues while maintaining a balance between stability and plasticity in dynamic environments. In contrast, artificial perception models deployed on resource-constrained edge devices are typically trained in a static offline manner and remain frozen after deployment, often suffering severe performance degradation under domain shifts. While large-scale models may encode broad knowledge through massive parameter redundancy, lightweight networks face a static optimization dilemma: forcing compact models to learn universal geometric representations is computationally inefficient and often leads to performance saturation. To resolve this issue, an Online Active Learning (OAL) mechanism is introduced to endow compact neural networks with the capability to adapt continuously during operation. A closed-loop Predict-Evaluate-Correct learning paradigm is established to actively select high-confidence, information-rich signals from streaming visual input. Crucially, Elastic Weight Consolidation (EWC) is employed not merely to prevent catastrophic forgetting, but to enforce Selective Plasticity, preserving parameters that encode globally relevant structural knowledge while allowing local alignment to newly observed environments. Built upon a MobileNetV3-Small backbone, the proposed system achieves approximately a 75% reduction in computational cost while maintaining competitive depth estimation accuracy. Experimental results demonstrate that adaptability is not solely determined by model size, but rather by how effectively parameter plasticity is regulated in dynamic environments.

Contemporary model-free reinforcement learning algorithms can achieve very high performance, but have low sample efficiency and are not robust to changes in the environment. Model-based algorithms have much higher sample efficiency, but still fail when the environment shifts. This paper introduces Adaptive Topological Learning with Abstract Successors (ATLAS) to combat these challenges. ATLAS uses a Grow When Required network with Successor Features in order to achieve high sample efficiency while also robustly tackling catastrophic forgetting. We evaluate ATLAS in spatial navigation tasks, benchmarking its performance against common on-policy and off-policy algorithms. Our empirical results demonstrate that by structurally decoupling transition dynamics from the reward signal, ATLAS achieves near-instantaneous adaptation to new goals and can exhibit positive backward transfer, significantly outperforming baseline methods in non-stationary environments.

Deploying robots in everyday human environments requires perception systems that are both robust and adaptable to diverse, dynamic conditions. In this work, we present a modular perception pipeline for household manipulation tasks, with a focus on dishware handling in kitchen environments. The pipeline integrates open-vocabulary object detection, multi-view segmentation, instance-aware 3D reconstruction, and a 2D-3D feature fusion strategy for 6D pose estimation and grasp planning. Its modular design enables systematic substitution of multiple visual and geometric foundation models, allowing us to identify the best-performing configuration through extensive evaluation on a custom kitchen dataset. The best-performing configuration (LLMDet + SAMv2 + DINOv2 + GeoTransformer) achieves an ADI of 89.12% on the 20-scene kitchen benchmark with cluttered and occluded conditions. Furthermore, real-world demonstrations confirm that the best configuration can be deployed on physical robots without environment-specific retraining, successfully executing tasks such as sink-to-dishwasher transfer and cup stacking. It validates the adaptability and scalability of the pipeline and highlights its potential as a practical framework for household robotic systems. Our code and supplementary materials are available at https://raivlab.github.io/FM_kitchen .

To operate robustly in open-world environments, autonomous agents should be able to infer the behavior of unfamiliar systems through interaction alone, even in the absence of documentation. However, existing tool-use benchmarks expose semantic tool schemas in static environments, allowing agents to rely on prior knowledge rather than autonomous discovery. To address this limitation, we introduce ScrambleToolBench, an interactive terminal benchmark designed to isolate behavioral reasoning. By removing semantic cues and enforcing a continuous task curriculum, the benchmark requires agents to uncover hidden tool behaviors entirely through trial-and-error interaction. The benchmark further introduces dynamic challenges, including mapping drift, stochastic action failures, and temporal execution windows, to evaluate whether agents can revise and adapt their hypotheses as the environment changes. Our evaluation of state-of-the-art language models reveals that successful initial discovery does not translate into robust adaptation. When faced with structural changes such as mapping drift, agents fail to use deductive strategies such as cycle tracing, and instead exhibit belief inertia or fall back to exhaustive search. Increasing test-time reasoning only amplifies this expensive brute-force search rather than enabling deductive recovery. While equipping agents with persistent memory reduces compounding errors, they remain unable to efficiently infer structural changes, highlighting a gap in current agent reasoning.

This paper extends the stochastic-oracle model of AI-augmented computing to include agentic oracles. Unlike a stationary stochastic oracle, which responds to the same query according to a fixed response distribution across calls, an agentic oracle can pursue a goal autonomously and may access an environment containing task-relevant resources. These capabilities affect both response distributions and token costs beyond what is visible at the query-response interface. We develop a framework for analyzing token costs in Stochastic-Oracle Turing Machines (SOTMs) that compute with agentic oracles. Each call has an \emph{orchestration token cost}, visible to the caller at the query-response interface, and an \emph{agentic token cost}, incurred by internal operations not exposed to the caller. We show that an SOTM computing with an agentic oracle that can retain intermediate state can have token-cost advantages over SOTMs using stationary stochastic oracles when solving the same task at the same quality level, both with and without environment access. We also investigate goal-loss risk, including how internal dispatch ordering can reduce exposure to irreversible actions. We provide a goal-loss avoidance criterion, derive progress--retry--goal-loss formulas, establish goal-depth lower bounds on token complexity, characterize token complexity when the probability of goal loss is zero, and show that goal-loss risk can impose an upper bound on the achievable quality of a task involving environment updates.

Configuration tuning is essential for optimizing the performance of highly configurable systems, e.g., throughput or runtime, under a given environment. Yet, this is a challenging process as there can be many options to tune, and configuration measurement is often highly expensive. In this paper, we demonstrate the phenomenon of ``less can be more'': system configuration tuning can be greatly improved with much superior budget utilization by partially tuning under the imperfect-fidelity---an environment that is similar, but cheaper to measure, compared with the concerned perfect-fidelity of environment under which the system should be tuned. We codify a conceptual framework of fidelity for configurable systems, drawing on which allows us to propose MFTune, a tuner that proactively explores in the space of $>10^4$ possible imperfect-fidelity settings to approximate a useful one, which strikes for the wideness of tuning. This creates high-quality seeds for the perfect-fidelity, which in turn ensures the tuning depth. Experiment results against $10$ state-of-the-art tuners, obtained from running diverse real-world systems for $19$ months $24 \times 7$, show that MFTune performs considerably better on $83.33$% cases with up to $19.34\%$ improvement while achieving hours of budget saving in general.

Physical-world interaction is inherently dynamic, as environments can evolve during execution, requiring agents to adapt their plans under non-stationary conditions. We study this challenge through long-horizon embodied planning under environment deviations and execution uncertainty. Existing embodied-task benchmarks can expose such failures, but these failures are usually treated as evaluation outcomes instead of learnable signals for training agents to recover. In this work, we introduce DynamicEnvPlan, a closed-loop framework for high-level planning in dynamic environments. It extends embodied task execution with humanoid agents, high-level primitive skills, structured semantic memory, and controllable perturbations. Our data synthesis design consists of planning, perturbation, and guarded correction modules that turn dynamic execution states into recovery-oriented traces. The resulting traces are used for staged supervised fine-tuning, enabling the planner to learn from both nominal execution and perturbed recovery trajectories. Using 104 task-scene combinations spanning i.i.d., compositional generalization, and out-of-distribution settings for fine-tuning and evaluation, DynamicEnvPlan boosts success rate from 33.3% for the base planner to 76.2%, while improving across all seven evaluation metrics critical to physical-world interaction, including safety and affordance compliance.

Graphical user interface (GUI) agents based on large language models are increasingly deployed across mobile, web, and desktop environments. However, existing agents are typically domain-specific, limiting the deployment and user experience. This motivates the consolidation of specialized models into a single cross-environment policy. Weight merging directly merges domain-specific experts but can corrupt executable actions under expert disagreement, while on-policy distillation (OPD) avoids conflicting teacher supervision yet still treats all response tokens equally during distillation, ignoring that action tokens are the only interface between the environment and the agent. To address this, We introduce MAGA that re-allocates training signal according to the structured action. Based on the correctness of the generated action, it suppresses unnecessary or invalid distillation signals and focuses learning on erroneous actions. Besides, a training-only hint optimizes the supervision signal provided by domain-specific teachers without changing the student input. Across two model scales, MAGA achieves the highest mean success rate, outperforming the strongest baseline by 2.0% at 8B and achieves almost the same average performance with teachers.

High-fidelity radio fields are typically simulated for every scene--transmitter configuration or fitted separately to each scene, failing to exploit propagation structures shared across environments. We present Point2Radio, a foundation model that learns a transferable propagation prior from multiple environments. Given a material-aware point cloud and a transmitter (TX) setting, a common encoder produces a TX-conditioned scene representation that can be queried at arbitrary receiver (RX) locations. Task-specific query decoders map this representation to different radio quantities, e.g., three-dimensional (3D) path-gain (PG) fields and power angular spectra (PAS). At inference for a new scene, the model uses only a material-aware point cloud and transceiver queries, running in milliseconds on a single GPU without meshes or explicit path tracing. We evaluate PG prediction on a scene-disjoint split of a 337-scene corpus containing 86,272 TX-conditioned fields. Point2Radio achieves 0.871 dB mean absolute error (MAE), reducing error by 76.7% relative to a same-split UNet-style baseline. The same encoder also supports PAS prediction via a task-specific decoder. Experiments further show that light target-scene fine-tuning improves adaptation to a specific environment.

Autonomous Cyber Operations (ACO) are increasingly important for defending enterprise networks as cyber threats continue to evolve in sophistication. ACO applications commonly employ Reinforcement Learning (RL) agents to learn defensive behaviors through interaction with environments. However, RL agents typically require extensive exploration during training, often resulting in unstable behavior and poor initial decision-making before converging toward effective defense strategies. In this work, we investigate the use of a Large Language Model (LLM) to improve autonomous defensive decision-making within an ACO environment. Through prompt engineering rather than fine-tuning, we demonstrate that an 8-billion parameter LLM pretrained on cybersecurity data can outperform a baseline RL agent in a modified CybORG CAGE Challenge 2 environment. We then propose an online policy distillation framework that transfers the LLM's defensive policy into a lightweight RL agent containing only 64,910 parameters, reducing model size by several orders of magnitude while maintaining effective defensive capabilities. This provides a pathway toward operationalizing frontier cybersecurity models within lightweight, deployable agents. To evaluate transferability, we construct CybORG scenarios ranging from 4 to 12 hosts and assess the approach across varying network configurations. We also evaluate teacher-guided RL stabilization strategies and observe that none consistently surpass the optimized teacher policy, suggesting policy-alignment limitations between reward-driven RL optimization and teacher-guided defense strategies. Our results demonstrate the potential of cybersecurity-focused LLMs as sources of expertise for autonomous cyber defense, while policy distillation provides a practical path toward operationalizing frontier cybersecurity models within efficient, scalable agents.

Existing evaluations often reduce agent failures to system-level outcomes, obscuring where the fault originated and which intervention would improve the agent system. This creates a repair-assignment problem: the same visible failure may call for model post-training, harness engineering, environment redesign, or benchmark repair depending on its source. Because agent behavior emerges from interactions among models, harnesses, users, tools, memory, and environments, outcome-level labels are often insufficient for improvement. Most failure taxonomies do little to resolve this problem because they are benchmark-specific and lack a shared structure. We introduce an interaction-centric taxonomy that localizes failures to the interactions in which they originate and identifies the responsible component. It organizes 41 failure modes by assigning each to an edge between two components and a fault side indicating where the repair belongs. This makes the taxonomy actionable: model-side failures identify targets for post-training, harness-side failures point to scaffolding and tool-integration fixes, and environment or grader failures reveal evaluation conditions requiring redesign. The schema applies across agent architectures, from coding assistants to long-horizon personal assistants and multi-agent systems. We ground the taxonomy in worked examples from public benchmarks, model system cards, published reports, and logged agent trajectories, and evaluate its reproducibility using independent reasoning agents as judges. Across four frontier models, the strongest judge reaches Cohen's $κ=0.76$ against human category labels, suggesting that the categories capture shared structure rather than annotator-specific preferences.

A central premise of latent world models is that predicting the future forces a representation to internalize the physics of its environment. Which physical quantities does a trained latent actually contain, and what decides this? We answer with controlled interventions in POKEWORLD, an interactive environment whose visually identical objects hide mass, drag, and contact stiffness. A certificate-gated protocol first certifies each parameter as recoverable from raw observations, then measures whether it enters the latent, so a null result can be attributed to the objective rather than to the environment. The resulting identifiability map has two organizing mechanisms and one frontier. Inputs limit what can be known, while prediction targets decide what is retained. Stiffness enters the latent only when touch is forecast ($R^2=0.50$, compared with $-0.02$ when the same signal is merely fused into the input), and under single-step prediction a vision-only latent discards even perfectly visible object state. Drag marks the frontier. It carries a recoverability certificate of 0.89 yet plateaus near 0.13 under every deterministic prediction objective we test, while a supervised head on the same trunk reaches 0.45. Parameters whose readout is slow and ratio-type under the sensed coordinates fall outside what these objectives acquire. On RH20T, an input-target factorial across scaling curves reproduces both mechanisms across two robots and 4,258 episodes. Every arm missing information or prediction pressure stays flat over a fivefold data range, and only the full multimodal objective forecasts force beyond a persistence baseline, with held-out gains that grow with scale. Objective structure determines which physical parameters a latent acquires, and additional data improves only the parameters it already acquires.